Nvidia

Llama-3.3-Nemotron-Super-49B-v1.5 is a 49B-parameter, English-centric reasoning/chat model derived from Meta’s Llama-3.3-70B-Instruct with a 128K context. It’s post-trained for agentic workflows (RAG, tool calling) via SFT across math, code, science, and multi-turn chat, followed by multiple RL stages; Reward-aware Preference Optimization (RPO) for alignment, RL with Verifiable Rewards (RLVR) for step-wise reasoning, and iterative DPO to refine tool-use behavior. A distillation-driven Neural Architecture Search (“Puzzle”) replaces some attention blocks and varies FFN widths to shrink memory footprint and improve throughput, enabling single-GPU (H100/H200) deployment while preserving instruction following and CoT quality. In internal evaluations (NeMo-Skills, up to 16 runs, temp = 0.6, top_p = 0.95), the model reports strong reasoning/coding results, e.g., MATH500 pass@1 = 97.4, AIME-2024 = 87.5, AIME-2025 = 82.71, GPQA = 71.97, LiveCodeBench (24.10–25.02) = 73.58, and MMLU-Pro (CoT) = 79.53. The model targets practical inference efficiency (high tokens/s, reduced VRAM) with Transformers/vLLM support and explicit “reasoning on/off” modes (chat-first defaults, greedy recommended when disabled). Suitable for building agents, assistants, and long-context retrieval systems where balanced accuracy-to-cost and reliable tool use matter.

NVIDIA-Nemotron-Nano-9B-v2 is a large language model (LLM) trained from scratch by NVIDIA, and designed as a unified model for both reasoning and non-reasoning tasks. It responds to user queries and tasks by first generating a reasoning trace and then concluding with a final response. The model's reasoning capabilities can be controlled via a system prompt. If the user prefers the model to provide its final answer without intermediate reasoning traces, it can be configured to do so.

Llama-3.1-Nemotron-Nano-8B-v1 is a compact large language model (LLM) derived from Meta's Llama-3.1-8B-Instruct, specifically optimized for reasoning tasks, conversational interactions, retrieval-augmented generation (RAG), and tool-calling applications. It balances accuracy and efficiency, fitting comfortably onto a single consumer-grade RTX GPU for local deployment. The model supports extended context lengths of up to 128K tokens. Note: you must include detailed thinking on in the system prompt to enable reasoning. Please see Usage Recommendations for more.

Llama-3.3-Nemotron-Super-49B-v1 is a large language model (LLM) optimized for advanced reasoning, conversational interactions, retrieval-augmented generation (RAG), and tool-calling tasks. Derived from Meta's Llama-3.3-70B-Instruct, it employs a Neural Architecture Search (NAS) approach, significantly enhancing efficiency and reducing memory requirements. This allows the model to support a context length of up to 128K tokens and fit efficiently on single high-performance GPUs, such as NVIDIA H200. Note: you must include detailed thinking on in the system prompt to enable reasoning. Please see Usage Recommendations for more.

Llama-3.1-Nemotron-Ultra-253B-v1 is a large language model (LLM) optimized for advanced reasoning, human-interactive chat, retrieval-augmented generation (RAG), and tool-calling tasks. Derived from Meta’s Llama-3.1-405B-Instruct, it has been significantly customized using Neural Architecture Search (NAS), resulting in enhanced efficiency, reduced memory usage, and improved inference latency. The model supports a context length of up to 128K tokens and can operate efficiently on an 8x NVIDIA H100 node. Note: you must include detailed thinking on in the system prompt to enable reasoning. Please see Usage Recommendations for more.

NVIDIA's Llama 3.1 Nemotron 70B is a language model designed for generating precise and useful responses. Leveraging Llama 3.1 70B architecture and Reinforcement Learning from Human Feedback (RLHF), it excels in automatic alignment benchmarks. This model is tailored for applications requiring high accuracy in helpfulness and response generation, suitable for diverse user queries across multiple domains. Usage of this model is subject to Meta's Acceptable Use Policy.

Nemotron-4-340B-Instruct is an English-language chat model optimized for synthetic data generation. This large language model (LLM) is a fine-tuned version of Nemotron-4-340B-Base, designed for single and multi-turn chat use-cases with a 4,096 token context length. The base model was pre-trained on 9 trillion tokens from diverse English texts, 50+ natural languages, and 40+ coding languages. The instruct model underwent additional alignment steps: 1. Supervised Fine-tuning (SFT) 2. Direct Preference Optimization (DPO) 3. Reward-aware Preference Optimization (RPO) The alignment process used approximately 20K human-annotated samples, while 98% of the data for fine-tuning was synthetically generated. Detailed information about the synthetic data generation pipeline is available in the technical report.